How to tell if text or an image was made by AI
A reader forwarded us a marketplace listing last month for a “near new” couch, photographed in a lounge room that had seven-fingered cushions and a window looking onto what was unmistakably nothing. That one was easy. Most aren’t. The honest answer to whether you can reliably tell if something was made by AI is: less and less, and the tools that promise to tell you are often worse than guessing.
That’s the uncomfortable starting point for this piece. We’re not going to pretend there’s a reliable detector app you can download and trust. There isn’t, not yet, and arguably not ever in the way people hope. What we can offer is a set of habits, the kind of scepticism that’s useful at the pub as much as online, for reading text and images with your eyes rather than outsourcing the judgement to software that claims certainty it doesn’t have.
Why detector tools don’t work the way people assume
Every few months a new “AI detector” does the rounds, promising a percentage score for how likely a piece of text is machine-written. We’d treat those scores the way we’d treat a free skin cancer check at a shopping centre kiosk: a rough prompt to look closer, not a diagnosis.
The fundamental problem is that these tools work by spotting statistical patterns common in AI writing, and large language models are explicitly being tuned to avoid those patterns. It’s an arms race, and the detector is almost always a step behind the model it’s trying to catch. Academic researchers have also found these tools produce more false positives on text written by non-native English speakers, which is a genuinely unfair outcome if a tool like that is used to fail a student or reject a job application.
Our take: if a tool’s business model depends on a confident-sounding score, be wary of the confidence more than the score.

Reading text for AI’s fingerprints
Generated text still tends to have a rhythm to it, even now. It often over-explains, restates the question in the opening line, and reaches for certain structures: three items in a list where one sentence would do, a tidy summary at the end that nobody asked for, phrases that sound encyclopaedic rather than personal.
Look for an absence of specific, checkable detail. Real human writing about, say, a local council decision will usually mention a date, a name, an actual address or meeting location. AI-generated text, especially when it’s hallucinating to fill gaps, tends to stay comfortably vague: “residents expressed concern” rather than naming who, where or when. If a piece of writing could have been written about any suburb in any state without changing a word, that’s worth noticing.
Email scams are the practical end of this. Older phishing emails were riddled with grammar mistakes that acted as an unintentional filter, useful to us because they were so easy to spot. AI has mostly fixed that problem for scammers. Scamwatch, run by the ACCC’s Scamwatch, has flagged that scam messages are getting harder to pick on spelling alone, which means the structural tells (urgency, generic greetings, mismatched sender domains) matter more than they used to. We’ve written before about why securing your email account matters more than almost any other single step, partly because a compromised inbox is now the launchpad for AI-polished scam messages sent “from” people you trust.
What still gives AI images away
Image generators have improved faster than most people realise, which is exactly the problem. Eighteen months ago the tells were obvious: mangled hands, nonsensical text on signage, ears that didn’t match. Some of that still shows up, particularly in crowd scenes or anything with small text in the background, but the big generators have quietly fixed a lot of it.
What’s harder to fix is context. Look at reflections and shadows, they often don’t agree with each other on where the light source is. Look at repeating patterns, like brick work or fabric weave, which can blur or repeat in odd ways the eye catches before the brain names it. Background objects at the edge of frame are still a weak point: a doorframe that doesn’t quite meet the wall, jewellery that shows an extra link, text on a distant shopfront that’s almost readable but isn’t actual words.
The other giveaway is less visual and more circumstantial. Ask where else this photo exists. A real photo of a real event usually has other angles, other photographers, a verifiable location. An AI image usually exists in exactly one place, posted by exactly one account, with no other version anywhere.
A short, dry truth: the fakes aren’t getting easier to spot technically, they’re getting easier to spot by asking who benefits from you believing it.
Watermarks and provenance tools, and their limits
Some major AI image tools now embed invisible metadata, sometimes called a content credential or watermark, that’s meant to travel with the file and identify it as AI-generated. Google’s SynthID and the C2PA content provenance standard, backed by a coalition that includes Adobe, Microsoft and others, are the two most talked-about approaches.
These are genuinely useful when they survive, which is the catch. Take a screenshot, crop the image, run it through a messaging app’s compression, or re-upload it to a platform that strips metadata, and the watermark can vanish before it reaches the person who needed to see it. We reckon these standards are worth having, the way seatbelts are worth having even though they don’t prevent every crash, but nobody should treat an absent watermark as proof an image is real. Absence of evidence isn’t evidence of absence here.
A mildly contrarian view on all this
Here’s where we’ll push back on the general mood. The popular framing is that we’re approaching a crisis where nobody will be able to trust anything they see online. We think that’s overstated, and slightly lazy. Photos have been manipulated since the darkroom existed, and entire industries, from newsrooms to courts to insurance assessors, built verification habits around that fact long before generative AI existed. The skills are the same ones: check the source, check for corroboration, be suspicious of anything designed to make you feel something fast. AI has lowered the cost of producing fakes, not invented the problem of fakes existing. The panic is doing more work than the actual shift in risk justifies for most ordinary browsing.
Where this actually bites in Australia right now
The practical risk for most readers isn’t being fooled by an arty AI image on social media. It’s AI-polished scam listings on marketplace sites, fake “before and after” renovation photos used to sell dodgy trade services, and cloned voice or video used in investment pitches. We’ve covered the broader pattern in our piece on investment scams on social media, and the warning signs there (urgency, guaranteed returns, a public figure “endorsing” something they’ve never heard of) now frequently come wrapped in AI-generated video or images that look far more polished than the old scam templates ever did.
There’s also a workplace angle worth a mention. AI-written reports, AI-generated headshots on fake LinkedIn profiles, and AI-summarised documents are now common enough in Australian offices that we think it’s worth reading our explainer on how to summarise documents with AI safely, mostly so you understand what the tool is doing to your own material and can apply the same scepticism to someone else’s.
If you do want to report something that looks like an AI-assisted scam, ReportCyber via the Australian Cyber Security Centre is the right first stop for anything involving financial loss or compromised accounts, separate from Scamwatch, which tracks the broader scam landscape.
The checks we actually bother with
Three things, done consistently, cover most situations:
- Search for the exact phrase or image elsewhere before reacting, because something real and newsworthy almost always has more than one source.
- Look at edges, reflections and small background text before you look at the main subject, since generators still spend less effort on the periphery.
- Treat emotional urgency, not technical polish, as the real red flag. Scammers used to rely on bad grammar to filter out careful readers; now they rely on panic to stop you reading carefully at all.
None of that is a detector tool. It’s closer to the habits a newsroom sub-editor had in 1995, checking a wire photo against a second source before running it. The technology changed, the discipline didn’t.
Our verdict: trust the habit, not the badge
We’d rather Australians develop a mild, permanent scepticism toward anything designed to be shared quickly than rely on a percentage score from a detector tool that’s wrong as often as it’s right. Provenance standards like C2PA are a good direction and worth supporting, but they’re a seatbelt, not a guarantee. The eSafety Commissioner publishes practical guidance for spotting manipulated media aimed at families, and it’s a reasonable place to send a relative who’s just forwarded you something that looks a bit too convenient.
The question “was this made by AI” is slowly becoming the wrong question. The better one, and the one that’s held up for decades before any of this existed, is simply: can I verify this somewhere else? If the answer is no, treat it as unverified, whoever or whatever made it.
— Dale Whitfield, Founding Editor, Tech Geek



